Back to Lab
Research Pillar

State Coherence & Persistence

Structured state stores with formal contracts ensuring deterministic reliability. Without coherent state, generated software would be dream-like — vivid but inconsistent.

Current Frontier

Formal state contracts: typed, versioned state schemas that the world model must respect.

Key Questions

01

Can a single world model maintain coherent state across arbitrarily complex interactions?

02

How do you detect and recover from state drift in continuous generation?

Key Papers

Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion

Chen et al.

Variable-horizon coherence without catastrophic drift

Read paper

PERSIST: Persistent 3D World Generation

March 2026

Persistent 3D state in latent space, camera as query, geometric consistency by construction. THE most important paper for state coherence.

Read paper

StateSpaceDiffuser

NeurIPS 2025

+14dB PSNR over DIAMOND, 50-step coherence vs 4-step via dual SSM+diffusion architecture.

Read paper

EDELINE: SSM World Models

NeurIPS 2025

Replaces DIAMOND's LSTM with Mamba SSM for immediate coherence gains.

Read paper

Self Forcing

NeurIPS 2025 Spotlight

Solves autoregressive drift by training on self-generated outputs. Directly attacks state drift.

Read paper

MultiGen: Memory for Diffusion Game Engines

March 2026

External memory for diffusion game engines. Decomposes into Memory + Observation + Dynamics modules.

Read paper

Long-Context SSM Video World Models

ICCV 2025

Block-wise SSM extends temporal memory beyond attention windows.

Read paper

DreamerV3

Nature 2025

General world model maintaining state across 150+ diverse tasks with single config.

Read paper

Current Insights

Separating state management from generation may be the key. The world model reads from and writes to a canonical state store, rather than maintaining state implicitly.

State contracts act as guardrails — the model can generate freely within the constraints of the contract, ensuring consistency.